topk(k, <bare selector>) by (labels) mislowers: implicit Sum eats the partition labels

[zzylol]

Summary

topk(k, <bare instant selector>) by (labels) — e.g. topk(3, http_requests_total) by (job) ("top-3 series per job") — does not lower correctly. This is a regression surfaced reviewing PR #18 (which removed the L3 Partition node and routed per-group ranking onto Sort.partition_by).

Repro

topk(3, http_requests_total) by (job)
→ column resolution failed: column `job` not found in schema (have: ["sum"])

Root cause

In crates/lower/src/promql.rs, build's Outer::TopK else-branch (the non-heavy-hitter path) defaults a bare selector argument (inner.func == None) to an implicit cross-series AggFunc::Sum:

let func = match &inner.func {
    Some(f) => inner_func(f),
    None => AggFunc::Sum,   // ← here
};
let base = windowed_aggregate(inner, vec![], func);

That reducing Sum collapses every label (including job) into a single sum column, so when the converter then resolves Sort.partition_by = [job] positionally against the child schema (lower.rs), job no longer exists → resolution error.

It is also semantically wrong independent of the by failure: PromQL topk(3, http_requests_total) ranks the raw instant-vector samples — it does not sum them.

Expected L3 shape

A bare selector should rank its own (label-preserving) samples:

topk(3, http_requests_total) by (job)
→ Limit { n: 3 } → Sort { keys: [value desc], partition_by: [job] } → Scan(http_requests_total)

This is exactly the shape PR #18's Sort.partition_by was designed for — no Partition node needed. Range-vector-function arguments (topk(3, rate(m[5m])) by (job), topk(3, avg_over_time(m[5m])) by (job)) already work because they produce a label-preserving per-series reduction.

Fix

In the Outer::TopK else-branch, when the argument is a bare instant selector, rank over the filtered_source directly (label-preserving) instead of wrapping it in a Sum. Only a range-vector-function argument gets the per-series reduction.

Tests

• topk(3, http_requests_total) by (job) → Limit → Sort{value desc, partition_by:[job]} → Scan (job preserved).
• topk(3, http_requests_total) (no by) → ranks raw samples, no implicit Sum.
• Regression: topk(k, count_over_time(...)) stays heavy-hitter; bottomk(...) stays generic Sort+Limit.

Follow-up to PR #18.